动态环境下解空间自学习多目标进化算法及应用研究

With the continuous emergence of dynamic complex problems in real applications, how to solve the large scale dynamic optimization problems has become an important research issue on evolutionary algorithm. This project will focus on a self-learning based multiobjective evolutionary algorithm on solution space under dynamic environments, and try to solve the drawbacks of the current algorithms on tracking dynamic optimal solutions and guiding evolutionary strategy design. Specifically, several works will be done for the deployment of this algorithm. A new model for complex dynamic solution space which based on the system dynamic algorithm has been set up and a dimension reduction algorithm is to be studied based on the Bayesian model and sparse subspace learning method; a fast learning algorithm is to be proposed based on an unsupervised online multi-kernel learning method; a mechanism of adaptive evolutionary strategy deisgn is to be studied, where some dynamic evolutioanry operators are to be given based on the dynamic Pareto optimal set model and a hybrid search strategy is to be designed based on the reinforcement learning and supervised learning methods; a new performance measurement based on robustness and convergence is to be proposed for the dynamic multiobjective evolutionary algorithms; and a typical application system prototype on the online portfolio strategy optimization is to be developed in order to test the efficiency of the proposed algorithm which will provide the directions for improvement on practice. In this project, the new ideas of how to track the dynamic optimal solutions and design the evolutionary strategies will be proposed, which will further provide new theories and methodologies for the dynamic multiobjective optimization algorithms, and valuable decision support for online portfolio optimization.

随着实际应用中动态复杂问题的不断涌现，面向大规模动态优化问题的求解成为进化算法研究及应用中的一个重要课题。本项目针对传统多目标进化算法求解过程中动态最优解集难以跟踪、进化策略缺乏指导等不足，研究动态环境下解空间自学习多目标进化算法，具体包括：1）基于系统动力学理论的复杂动态解空间数据建模方法及基于贝叶斯模型和稀疏子空间学习的数据降维算法；2）基于在线多核学习理论的解空间结构学习算法；3）动态自适应进化策略设计与优化，包括基于最优解集结构概率模型的进化算子设计方法、基于强化学习和有监督学习的混合搜索策略设计与优化方法；4）基于鲁棒性和收敛性的动态多目标优化性能测度指标构造；5）以在线投资决策优化为典型应用，验证算法的有效性，并提出实用的改进方向。通过本项目的研究，为动态最优解集跟踪、进化策略设计提供新的思路，为动态多目标进化算法研究提供新的理论和方法，也为在线投资提供有参考价值的决策支持。

面向大规模动态优化问题的求解是目前进化算法研究及应用中的一个重要课题。本项目针对传统多目标进化算法求解过程中动态最优解集难以跟踪、进化策略缺乏指导等不足，研究动态环境下解空间自学习多目标进化算法，通过运用机器学习的理论方法对复杂解空间进行建模分析和学习预测，提出高效的动态多目标优化算法，提高算法对复杂问题的求解效率。在复杂解空间建模、解空间学习、算法与搜索策略设计、算法应用推广等方面取得了若干具有创新性的成果，实现了项目预期研究目标。培养了博士生2人，硕士生7人，发表学术论文26篇（包括期刊论文22篇，会议论文4篇），其中SCI 1区论文10篇，有3篇论文入选ESI高被引论文，授权专利1项。